Active Basis

1
Active Basis for Modeling, Learning and
Recognizing Deformable Template Ying Nian
Wu UCLA Department of Statistics Joint work with
Zhangzhang Si, Haifeng Gong, and Song Chun
Zhu Summer 2009
2
Reproducibility page http//www.stat.ucla.edu/ywu
/ActiveBasis Matlab/C code, Data
3

• Summary
• Representation Active basis model
• Learning Shared sketch algorithm
• Inference Sum-max maps

4
Gabor wavelets Daugman, 1985
Olshausen,
Field, 1996
Localized sine and cosine waves
Model for simple cells in primary visual cortex
5
Gabor wavelets
Operation local Fourier transform, edge detection
Representation wavelet sparse coding, Olshausen,
Field, 1996
raw intensities ? composition of strokes
6
Active basis

• A select set of Gabor wavelet elements
• Each element can perturb its location and
  orientation
• These elements form a deformable template

7
Active basis
An active basis can deform to sketch multiple
instances
8
(No Transcript)
9
Shared sketch algorithm
10
Shared sketch algorithm

• Decreasing order in log-likelihood ratio
  (template matching score)
• Background edges ignored
• First experiment tried

11
(No Transcript)
12
(No Transcript)
13
Recognizing learned deformable template
SUM1 Is there an edge here? simple cells
Daugman, 1985 MAX1 Is there an edge nearby?
complex cells local maximum pooling,
Riesenhuber and Poggio, 1999 SUM2 Is there a
composite sketch of edges here?
shape filter for template matching Soft scoring
instead of hard decision
14
SUM-MAX maps
15
Template matching by SUM-MAX
SUM2 map at optimal resolution
Multiple resolutions
16
(No Transcript)
17
(No Transcript)
18
(No Transcript)
19
Geometric transformation
Scaling, rotation, change of aspect ratio
20
(No Transcript)
21
Classification
Freund and Schapire, 1995 Viola and Jones, 2004
22
(No Transcript)
23
(No Transcript)
24
(No Transcript)
25
Learning from non-aligned training images
26
Learning from non-aligned training images

• Given the bounding box of one training image
• Can be relaxed by multiple starting

Motif finding
27
No given bounding box of any training image
28
Weizmann horse images
29
(No Transcript)
30
(No Transcript)
31
(No Transcript)
32
(No Transcript)
33
(No Transcript)
34
(No Transcript)
35
(No Transcript)
36
(No Transcript)
37
(No Transcript)
38
(No Transcript)
39
Learning part templates or visual words
40
Learning moving template from video sequence
PETS data set
41
EM/K-mean Clustering
42
EM/K-mean Clustering
43
EM/K-mean Clustering
44
Learning local representatives
Automatically finding local clusters in the image
ensemble and learn local representative
templates Local dimension reduction and local
metric
45
(No Transcript)
46
eps
eps
eps
eps
eps
eps
eps
47
(No Transcript)
48
MNIST data set
49
Including Weizmann data set and INRIA data set
50
Synthesis by multi-sale Gabors and DoGs
Low frequency Gabors and DoGs capture region
properties
51
(No Transcript)
52
Composing multiple templates
Learn bike template
Split bike template to detect and sketch tandem
bike
53
Is there a tandem bike here?
Is there a wheel nearby?
Is there a wheel here?
Is there an edge nearby?
Is there an edge here?
Soft scoring instead of hard decision
54
Where to split the bike template?
Define and learn parts as highly alignable
sub-templates
55
Large deformations
Parts to account for correlated activities
56

• Summary
• Model Active basis model
• Learning Shared sketch algorithm
• Inference Sum-max maps
• Simplicity
• Model wavelet expansion perturbation
• Learning edge detection parallel scoring
• Inference Gabor local max sum filtering
• Unsupervised supervised detection single
  image learning

57

• Principles
• Sparsity Olshausen-Field
• Compositionality S. Geman Zhu-Mumford
• Invariance Riesenhuber-Poggio
• Acknowlegement
• Alan Yuille, Zhuowen Tu, Leo Zhu, Chuck Fleming
• NSF-DMS 0707055, NSF-IIS 0713652
• Lotus Hill Institute

58

• Appendix Theoretical underpinnings
• shared matching pursuit
• regimes of patterns

59
Matching pursuit
Mallat, Zhang, 1993
single image
Step 3 explain-away local inhibition
a selected element inhibits highly correlated
ones hard inhibition inhibits those
overlapping in both spatial and frequency
domains

60
Shared matching pursuit
multiple images
Step 2 local maximization, perturb the active
element to sketch a nearby edge Step 3 local
inhibition
61
Shared matching pursuit
Pool statistics over parallel regressions
Approximation Step 3 hard inhibition,
non-overlapping
maximum contrast
62
Pursuit index and template matching score

• Vote for selecting the next element
• h-function monotone increasing
• discounts strong edges in
  background

Log-likelihood scoring Active correlation linear
scoring
non-probabilistic
63
Background density
64
FRAME --- Filters, Random field, And Maximum
Entropy
Zhu,Wu,Mumford,97 Wu,Zhu,Liu,00
65
Textures
Zhu,Liu,Wu,00 Wu,Zhu,Liu,00
66
Generic objects line segments, junctions, etc.
67
Guo, Zhu, Wu, 03,07
68
Scaling connection
Wu, Guo, Zhu, 2008
fine
coarse
Central limit theorem Entropy rate increases